pith. sign in

arxiv: 2605.17188 · v1 · pith:UAY56RLRnew · submitted 2026-05-16 · 📡 eess.IV

RDDM: A Residual-Driven Drifting Model for High-Fidelity Low-Dose CT Denoising

Jianxu Wang , Qing Lyu , Ge Wang This is my paper

Pith reviewed 2026-05-20 13:47 UTC · model grok-4.3

classification 📡 eess.IV
keywords low-dose CT denoisingresidual drifting modelgenerative modelsone-step inferencemedical image reconstructiondrifting modelsCT denoisingreal-time imaging
0
0 comments X

The pith

RDDM uses residual drifting forces to achieve one-step high-fidelity LDCT denoising at real-time speeds.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces RDDM to overcome the limitations of existing LDCT denoising approaches, where pixel-level methods produce overly smooth results and generative diffusion models deliver better fidelity but at the cost of slow multi-step inference. RDDM incorporates multi-step distribution evolution directly into training by defining a residual drifting field that applies attractive forces from the differences between low-dose and normal-dose CT images alongside repulsive forces from its own generated residuals. This design allows the model to learn faithful one-step denoising without post-processing or loss of anatomical detail. The authors demonstrate that this yields state-of-the-art performance on supervised baselines, including top quantitative scores and realistic textures, while running in milliseconds per slice. A reader would care because it points toward practical clinical tools that reduce radiation exposure without compromising diagnostic image quality or workflow speed.

Core claim

RDDM incorporates the multi-step distribution evolution from drifting models into the training dynamics through a residual drifting field formed by an attractive force induced by the residuals between LDCT and NDCT and a repulsive force induced by the generated residuals, thereby enabling one-step denoising. By adjusting parameter settings and adding pixel-level supervision, three variants are created to span needs from detail preservation to stronger noise suppression. Experiments show RDDM achieves state-of-the-art denoising among supervised baselines, with RDDM-Fine producing reconstructions highly consistent with NDCT, superior PSNR and SSIM, the best FID of 5.87, and realistic textures,

What carries the argument

The residual drifting field, which combines attractive forces from LDCT-NDCT residuals and repulsive forces from generated residuals to embed multi-step evolution into one-step training and inference.

If this is right

  • Three RDDM variants let users trade off between fine detail preservation and stronger noise removal by changing parameters and adding pixel supervision.
  • RDDM-Fine reaches the best reported FID of 5.87 together with high PSNR and SSIM while keeping realistic anatomical textures.
  • Denoising a 512 by 512 slice requires roughly 15 ms, supporting on-the-fly clinical use.
  • Reconstructions remain highly consistent with normal-dose CT references.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The force-based residual formulation could be tested on other low-signal medical imaging tasks such as MRI or PET denoising where iterative generative models are currently too slow.
  • Embedding attractive and repulsive residual dynamics might allow similar one-step approximations in non-medical generative modeling domains.
  • Real-time capability opens the possibility of embedding the model inside scanner reconstruction pipelines to enable immediate low-dose protocols.

Load-bearing premise

The multi-step distribution evolution of drifting models can be captured during training by a single residual drifting field of attractive and repulsive forces.

What would settle it

A controlled test on held-out LDCT data where the one-step RDDM output is compared directly to multi-step diffusion outputs on the same inputs and shows measurably lower structural fidelity or higher artifact rates than the iterative baseline.

Figures

Figures reproduced from arXiv: 2605.17188 by Ge Wang, Jianxu Wang, Qing Lyu.

Figure 1
Figure 1. Figure 1: Overview of the proposed RDDM framework. Top: the RDDM training framework, where a single residual sample is generated by the denoiser, while [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Residual power spectrum (RPS) visualization of LDCT and different [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison of denoising results from different methods on a quarter-dose sample from the Mayo dataset. The display window is [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Normalized radial 1D noise power spectrum (NPS) profiles averaged [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
read the original abstract

Low-dose CT (LDCT) denoising remains an important yet challenging problem in medical imaging. Although recent learning-based methods have shown promising performance, those optimized using classical pixel-level objectives often produce over-smoothed reconstructions. Existing mainstream generative models, such as diffusion models, have improved fidelity at the cost of expensive multi-step iterative inference, which limits their practicality for real-time use. To address this gap, we propose a Residual-Driven Drifting Model (RDDM) for effective, efficient, and high-fidelity LDCT denoising. Inspired by the recently proposed Drifting Models, RDDM incorporates the multi-step distribution evolution into the training dynamics through a residual drifting field, thereby enabling one-step denoising. Specifically, the residual drifting field is formed by an attractive force induced by the residuals between LDCT and normal-dose CT (NDCT) and a repulsive force induced by the generated residuals. In addition, by adjusting the parameter settings and incorporating pixel-level supervision, we develop three RDDM variants, covering application needs from detail preservation to stronger noise suppression. Extensive experiments demonstrate that RDDM achieves state-of-the-art denoising performance among supervised baselines. In particular, RDDM-Fine produces reconstructions that are highly consistent with NDCT, achieving superior PSNR and SSIM together with the best FID of 5.87 while preserving realistic anatomical textures. Moreover, RDDM enables on-the-fly inference, requiring only about 15 ms to denoise a single 512 x 512 LDCT slice. These results establish RDDM as a promising solution for high-fidelity and real-time LDCT denoising in clinical applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a Residual-Driven Drifting Model (RDDM) for high-fidelity low-dose CT (LDCT) denoising. Inspired by drifting models, RDDM incorporates multi-step distribution evolution into training via a residual drifting field formed by attractive forces from LDCT-NDCT residuals and repulsive forces from generated residuals, enabling one-step inference. Three variants are developed through parameter adjustments and pixel-level supervision. Experiments claim state-of-the-art performance among supervised baselines, with RDDM-Fine achieving superior PSNR/SSIM, the best FID of 5.87, realistic anatomical textures, and ~15 ms inference for 512x512 slices.

Significance. If the performance claims and the effectiveness of the residual drifting field for one-step inference hold, the work would be significant for medical imaging by addressing the fidelity-speed trade-off in LDCT denoising. It offers a practical alternative to multi-step generative models while preserving textures, with variants providing application flexibility. The fast inference time is a clear strength for clinical deployment. Credit is due for the novel residual-driven construction and the reported metrics, though these require stronger validation to fully realize the potential impact.

major comments (2)
  1. [Section 3 (residual drifting field definition)] The central claim that the residual drifting field (attractive forces from LDCT-NDCT residuals plus repulsive forces from generated residuals) captures multi-step distribution evolution to enable faithful one-step inference without mode collapse or bias lacks any derivation, fixed-point analysis, or convergence guarantee. This is load-bearing for the one-step fidelity assertion and the overall contribution.
  2. [Section 4 (experiments and results)] The state-of-the-art performance claim, including the FID of 5.87 for RDDM-Fine and superiority over supervised baselines, rests on limited evidence. The manuscript lacks details on datasets, specific baselines, ablation studies, and statistical testing, undermining assessment of robustness and generalizability.
minor comments (2)
  1. [Section 3.3] Clarify the exact parameter settings for the three RDDM variants and how they trade off detail preservation versus noise suppression.
  2. [Section 4] Add more context in tables or figures for the comparison methods and test conditions to support the reported PSNR, SSIM, and FID values.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback on our manuscript. We have carefully reviewed each major comment and provide point-by-point responses below, indicating where we will strengthen the submission through revisions.

read point-by-point responses

  1. Referee: [Section 3 (residual drifting field definition)] The central claim that the residual drifting field (attractive forces from LDCT-NDCT residuals plus repulsive forces from generated residuals) captures multi-step distribution evolution to enable faithful one-step inference without mode collapse or bias lacks any derivation, fixed-point analysis, or convergence guarantee. This is load-bearing for the one-step fidelity assertion and the overall contribution.

    Authors: We appreciate the referee highlighting the importance of theoretical grounding for the residual drifting field. The construction is motivated by the multi-step evolution in drifting models, where the attractive force pulls toward the LDCT-NDCT residual and the repulsive force prevents collapse by pushing against generated residuals. While the original manuscript presents this as an empirical design choice validated by results, we agree a more explicit discussion would help. In revision, we will expand Section 3 with additional motivation, a qualitative fixed-point analysis illustrating how the combined forces approximate the target distribution in one step, and further empirical observations from the training process showing stable behavior without mode collapse. A full convergence proof remains outside the current scope but the added analysis will better support the one-step inference claim. revision: yes

  2. Referee: [Section 4 (experiments and results)] The state-of-the-art performance claim, including the FID of 5.87 for RDDM-Fine and superiority over supervised baselines, rests on limited evidence. The manuscript lacks details on datasets, specific baselines, ablation studies, and statistical testing, undermining assessment of robustness and generalizability.

    Authors: We acknowledge that greater experimental transparency is needed to allow full assessment of the claims. The current manuscript reports results on established LDCT benchmarks and compares against several supervised methods, but we agree more detail is warranted. In the revised manuscript, we will expand Section 4 to provide: explicit dataset descriptions including acquisition parameters and split details; a complete enumeration of all supervised baselines with references and implementation notes; comprehensive ablation studies isolating the attractive/repulsive force components and the three RDDM variants; and statistical testing (means, standard deviations, and significance tests across repeated evaluations). These additions will strengthen the evidence for the reported PSNR/SSIM/FID gains and the method's robustness. revision: yes

Circularity Check

0 steps flagged

No circularity: RDDM is an empirical modeling proposal validated experimentally

full rationale

The paper defines RDDM explicitly as a new residual drifting field (attractive forces from LDCT-NDCT residuals plus repulsive forces from generated residuals) that is incorporated into training dynamics, with additional pixel-level supervision and parameter tuning to produce three variants. Performance is reported via standard metrics (PSNR, SSIM, FID) and timing on experimental data rather than any algebraic identity or self-referential fit. The central claim that this construction enables faithful one-step inference is presented as a modeling choice whose validity is tested empirically, not derived by construction from the inputs. No load-bearing step reduces to a self-citation chain, fitted parameter renamed as prediction, or definitional equivalence.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that drifting-model dynamics can be approximated by a residual field and on ad-hoc parameter choices for the three variants; no new physical entities are postulated.

free parameters (1)
  • parameter settings for RDDM variants
    Adjusted to span detail preservation to stronger noise suppression needs
axioms (1)
  • domain assumption Multi-step distribution evolution can be incorporated into training dynamics through a residual drifting field
    Directly stated as inspiration from recently proposed Drifting Models
invented entities (1)
  • residual drifting field no independent evidence
    purpose: To induce attractive force from LDCT-NDCT residuals and repulsive force from generated residuals for one-step denoising
    Newly defined construct that forms the core of the proposed model

pith-pipeline@v0.9.0 · 5825 in / 1351 out tokens · 49644 ms · 2026-05-20T13:47:46.419688+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · 6 internal anchors

  1. [1]

    Projected lifetime cancer risks from current computed tomography imaging,

    R. Smith-Bindman, P. W. Chu, H. Azman Firdaus, C. Stewart, M. Malekhedayat, S. Alber, W. E. Bolch, M. Mahendra, A. Berrington de Gonz´alez, and D. L. Miglioretti, “Projected lifetime cancer risks from current computed tomography imaging,”JAMA internal medicine, vol. 185, no. 6, pp. 710–719, 2025

    work page 2025

  2. [2]

    Computed tomography—an increasing source of radiation exposure,

    D. J. Brenner and E. J. Hall, “Computed tomography—an increasing source of radiation exposure,”New England journal of medicine, vol. 357, no. 22, pp. 2277–2284, 2007

    work page 2007

  3. [3]

    On gabor’s contribu- tion to image enhancement,

    M. Lindenbaum, M. Fischer, and A. Bruckstein, “On gabor’s contribu- tion to image enhancement,”Pattern recognition, vol. 27, no. 1, pp. 1–8, 1994

    work page 1994

  4. [4]

    Analysis of multiresolution image denoising schemes using generalized gaussian and complexity priors,

    P. Moulin and J. Liu, “Analysis of multiresolution image denoising schemes using generalized gaussian and complexity priors,”IEEE trans- actions on Information Theory, vol. 45, no. 3, pp. 909–919, 1999

    work page 1999

  5. [5]

    Image denoising by sparse 3-d transform-domain collaborative filtering,

    K. Dabov, A. Foi, V . Katkovnik, and K. Egiazarian, “Image denoising by sparse 3-d transform-domain collaborative filtering,”IEEE Transactions on image processing, vol. 16, no. 8, pp. 2080–2095, 2007

    work page 2080

  6. [6]

    Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization,

    E. Y . Sidky and X. Pan, “Image reconstruction in circular cone-beam computed tomography by constrained, total-variation minimization,” Physics in Medicine & Biology, vol. 53, no. 17, pp. 4777–4807, 2008

    work page 2008

  7. [7]

    Projection space denoising with bilateral filtering and ct noise modeling for dose reduction in ct,

    A. Manduca, L. Yu, J. D. Trzasko, N. Khaylova, J. M. Kofler, C. M. McCollough, and J. G. Fletcher, “Projection space denoising with bilateral filtering and ct noise modeling for dose reduction in ct,”Medical physics, vol. 36, no. 11, pp. 4911–4919, 2009

    work page 2009

  8. [8]

    Deep learning on image denoising: An overview,

    C. Tian, L. Fei, W. Zheng, Y . Xu, W. Zuo, and C.-W. Lin, “Deep learning on image denoising: An overview,”Neural Networks, vol. 131, pp. 251– 275, 2020

    work page 2020

  9. [9]

    A review on deep learning approaches for low-dose computed tomography restoration,

    K. S. H. Kulathilake, N. A. Abdullah, A. Q. M. Sabri, and K. W. Lai, “A review on deep learning approaches for low-dose computed tomography restoration,”Complex & Intelligent Systems, vol. 9, no. 3, pp. 2713– 2745, 2023

    work page 2023

  10. [10]

    Low-dose ct with a residual encoder-decoder convolutional neural network,

    H. Chen, Y . Zhang, M. K. Kalra, F. Lin, Y . Chen, P. Liao, J. Zhou, and G. Wang, “Low-dose ct with a residual encoder-decoder convolutional neural network,”IEEE transactions on medical imaging, vol. 36, no. 12, pp. 2524–2535, 2017

    work page 2017

  11. [11]

    Competitive performance of a modularized deep neural network compared to commercial algorithms for low-dose ct image reconstruction,

    H. Shan, A. Padole, F. Homayounieh, U. Kruger, R. D. Khera, C. Niti- warangkul, M. K. Kalra, and G. Wang, “Competitive performance of a modularized deep neural network compared to commercial algorithms for low-dose ct image reconstruction,”Nature Machine Intelligence, vol. 1, no. 6, pp. 269–276, 2019

    work page 2019

  12. [12]

    Ctformer: convolution-free token2token dilated vision transformer for low-dose ct denoising,

    D. Wang, F. Fan, Z. Wu, R. Liu, F. Wang, and H. Yu, “Ctformer: convolution-free token2token dilated vision transformer for low-dose ct denoising,”Physics in Medicine & Biology, vol. 68, no. 6, p. 065012, 2023

    work page 2023

  13. [13]

    Noise2Noise: Learning Image Restoration without Clean Data

    J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras, M. Aittala, and T. Aila, “Noise2noise: Learning image restoration without clean data,”arXiv preprint arXiv:1803.04189, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  14. [14]

    Iterative material decomposition for spectral ct using self-supervised noise2noise prior,

    W. Fang, D. Wu, K. Kim, M. K. Kalra, R. Singh, L. Li, and Q. Li, “Iterative material decomposition for spectral ct using self-supervised noise2noise prior,”Physics in medicine & biology, vol. 66, no. 15, p. 155013, 2021

    work page 2021

  15. [15]

    Low-dose ct reconstruction with noise2noise network and testing-time fine-tuning,

    D. Wu, K. Kim, and Q. Li, “Low-dose ct reconstruction with noise2noise network and testing-time fine-tuning,”Medical Physics, vol. 48, no. 12, pp. 7657–7672, 2021

    work page 2021

  16. [16]

    Deep embedding- attention-refinement for sparse-view ct reconstruction,

    W. Wu, X. Guo, Y . Chen, S. Wang, and J. Chen, “Deep embedding- attention-refinement for sparse-view ct reconstruction,”IEEE Transac- tions on Instrumentation and Measurement, vol. 72, pp. 1–11, 2022

    work page 2022

  17. [17]

    Noise2void-learning denoising from single noisy images,

    A. Krull, T.-O. Buchholz, and F. Jug, “Noise2void-learning denoising from single noisy images,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2129–2137

    work page 2019

  18. [18]

    Noise2self: Blind denoising by self- supervision,

    J. Batson and L. Royer, “Noise2self: Blind denoising by self- supervision,” inInternational conference on machine learning. PMLR, 2019, pp. 524–533

    work page 2019

  19. [19]

    Self2self with dropout: Learning self-supervised denoising from single image,

    Y . Quan, M. Chen, T. Pang, and H. Ji, “Self2self with dropout: Learning self-supervised denoising from single image,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 1890–1898. 9

    work page 2020

  20. [20]

    Noise suppression with similarity-based self-supervised deep learning,

    C. Niu, M. Li, F. Fan, W. Wu, X. Guo, Q. Lyu, and G. Wang, “Noise suppression with similarity-based self-supervised deep learning,”IEEE transactions on medical imaging, vol. 42, no. 6, pp. 1590–1602, 2022

    work page 2022

  21. [21]

    A fast blind zero-shot denoiser,

    J. Lequyer, R. Philip, A. Sharma, W.-H. Hsu, and L. Pelletier, “A fast blind zero-shot denoiser,”Nature Machine Intelligence, vol. 4, no. 11, pp. 953–963, 2022

    work page 2022

  22. [22]

    Zs-kan: Zero-shot image denoising with lightweight kolmogorov-arnold networks,

    J. Wang and G. Wang, “Zs-kan: Zero-shot image denoising with lightweight kolmogorov-arnold networks,”IEEE Transactions on Biomedical Engineering, 2025

    work page 2025

  23. [23]

    Zero-shot noise2noise: Efficient image de- noising without any data,

    Y . Mansour and R. Heckel, “Zero-shot noise2noise: Efficient image de- noising without any data,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14 018–14 027

    work page 2023

  24. [24]

    Median2median: Zero-shot suppression of structured noise in images,

    J. Wang and G. Wang, “Median2median: Zero-shot suppression of structured noise in images,”arXiv preprint arXiv:2510.01666, 2025

    work page arXiv 2025

  25. [25]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840– 6851, 2020

    work page 2020

  26. [26]

    Denoising Diffusion Implicit Models

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  27. [27]

    Poisson flow generative models,

    Y . Xu, Z. Liu, M. Tegmark, and T. Jaakkola, “Poisson flow generative models,”Advances in Neural Information Processing Systems, vol. 35, pp. 16 782–16 795, 2022

    work page 2022

  28. [28]

    Pfgm++: Unlocking the potential of physics-inspired generative models,

    Y . Xu, Z. Liu, Y . Tian, S. Tong, M. Tegmark, and T. Jaakkola, “Pfgm++: Unlocking the potential of physics-inspired generative models,” in International Conference on Machine Learning. PMLR, 2023, pp. 38 566–38 591

    work page 2023

  29. [29]

    Corediff: Contextual error-modulated generalized diffusion model for low-dose ct denoising and generalization,

    Q. Gao, Z. Li, J. Zhang, Y . Zhang, and H. Shan, “Corediff: Contextual error-modulated generalized diffusion model for low-dose ct denoising and generalization,”IEEE Transactions on Medical Imaging, vol. 43, no. 2, pp. 745–759, 2023

    work page 2023

  30. [30]

    Cocodiff: a contextual conditional diffusion model for low-dose ct image denoising,

    Q. Gao and H. Shan, “Cocodiff: a contextual conditional diffusion model for low-dose ct image denoising,” inDevelopments in x-ray tomography XIV, vol. 12242. SPIE, 2022, pp. 92–98

    work page 2022

  31. [31]

    Ppfm: Image denoising in photon-counting ct using single-step posterior sampling poisson flow generative models,

    D. Hein, S. Holmin, T. Szczykutowicz, J. S. Maltz, M. Danielsson, G. Wang, and M. Persson, “Ppfm: Image denoising in photon-counting ct using single-step posterior sampling poisson flow generative models,” IEEE Transactions on Radiation and Plasma Medical Sciences, vol. 8, no. 7, pp. 788–799, 2024

    work page 2024

  32. [32]

    Tomographic foundation model—force: Flow-oriented reconstruction conditioning engine,

    W. Xia, C. Niu, and G. Wang, “Tomographic foundation model—force: Flow-oriented reconstruction conditioning engine,”IEEE Transactions on Medical Imaging, 2026

    work page 2026

  33. [33]

    Generative Modeling via Drifting

    M. Deng, H. Li, T. Li, Y . Du, and K. He, “Generative modeling via drifting,”arXiv preprint arXiv:2602.04770, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  34. [34]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in International conference on machine learning. pmlr, 2015, pp. 2256– 2265

    work page 2015

  35. [35]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differ- ential equations,”arXiv preprint arXiv:2011.13456, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  36. [36]

    Flow Matching for Generative Modeling

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,”arXiv preprint arXiv:2210.02747, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  37. [37]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,”arXiv preprint arXiv:2209.03003, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  38. [38]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

    work page 2015

  39. [39]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    work page 2004